Most vulnerability programs do not fail from a shortage of CVEs. They fail from a shortage of judgment. When every finding is “urgent,” nothing is. CVSS alone answers how bad a flaw could be in a vacuum. It does not answer whether anyone is exploiting it this week, whether your asset is reachable, or whether the business will notice if that host falls over on a Tuesday.
Seasoned prioritization treats severity as one input among several — exploitation probability, known active exploitation, and business context — then produces a queue a remediation team can execute without debating philosophy in every standup.
The problem with CVSS-only triage
A CVSS 9.8 on an isolated lab box and a CVSS 9.8 on a production identity provider are not the same emergency. Teams that ignore that distinction either burn out patching the wrong things or normalize living with criticals until an incident makes the decision for them.
You need layered signals:
- CVSS — technical severity baseline
- EPSS — estimated probability of exploitation in the wild (FIRST.org)
- CISA KEV — confirmed active exploitation catalog
- Asset criticality & exposure — business impact and reachability
A composite model that stays explainable
Fancy machine learning is optional. Explainability is not — especially when auditors or engineering leads ask why ticket A beat ticket B:
risk_score = (cvss_norm × 0.30)
+ (epss × 0.30)
+ (kev_flag × 0.25)
+ (asset_criticality × 0.15)
Where kev_flag is 1.0 if the CVE is on CISA KEV, else 0. KEV membership should elevate priority even when EPSS has not caught up. Weights are starting points — tune them, document them, and stop pretending the defaults are sacred.
Data sources that keep the model honest
- NVD / CVE JSON — metadata and CVSS vectors
- EPSS API — daily probability scores
- CISA KEV feed — actively exploited vulnerabilities
- Asset inventory — CMDB or scanner tags for criticality and internet exposure
Garbage inventory produces garbage prioritization. Fix ownership and tiering before arguing about decimal places in the formula.
Workflow from scanner noise to remediation queue
- Ingest findings with CVE IDs from your scanners
- Enrich each CVE with CVSS, latest EPSS, and KEV membership
- Map to assets; apply criticality and exposure weight
- Compute composite score; sort descending
- Output a queue with evidence fields — score components included — for operators and auditors
Python enrichment sketch
def compute_risk(cvss, epss, in_kev, asset_tier):
cvss_norm = min(cvss / 10.0, 1.0)
kev_boost = 1.0 if in_kev else 0.0
crit_norm = asset_tier / 5.0
return round(
cvss_norm * 0.30
+ epss * 0.30
+ kev_boost * 0.25
+ crit_norm * 0.15,
3,
)
Keep the function boring. Boring functions get reviewed. Clever black boxes get ignored when production is on fire.
Operational rules that prevent debate loops
- Remediate KEV-listed CVEs on internet-facing assets within hours or days — not “next quarter”
- Use EPSS thresholds (for example > 0.5) as secondary triage for non-KEV items
- Re-score on a cadence — EPSS moves as threat intelligence moves
- Document the formula and exceptions for SOC 2 / ISO evidence
- Allow human override — with a written reason — for fragile systems and compensating controls
What this will not solve
No scoring model fixes missing owners, frozen change windows, or a culture that treats every exception as permanent. Prioritization is a lens. Remediation is still work. If leadership will not fund the work, the scoreboard only documents decline more precisely.
Closing thought
The goal is reproducible, evidence-backed prioritization — not subjective severity theater. Combine severity, exploitation likelihood, known exploitation, and business context. Then ship the queue, measure aging, and adjust weights when reality disagrees with your spreadsheet. That is how CVE volume becomes risk reduction instead of noise.
← Work & Write-ups